Responsibility Attribution in Human Interactions with Everyday AI Systems

Research Background and Issues

• Identified Problems or Challenges:
  The authors investigated how humans attribute responsibility in interactions with everyday artificial intelligence (AI) systems, particularly in scenarios involving joint decision-making between humans and AI. The study focuses on the impact of positive and negative outcomes as well as the observer and executor perspectives on responsibility attribution.

• Significance:
  As AI systems become increasingly prevalent, the complexity of responsibility attribution in decision-making involving AI becomes more challenging, especially in human-AI interactions. Understanding this issue can influence user trust in AI systems, ethical design, and decision-making outcomes.

• Research Motivation and Related Work:
  Previous studies have primarily focused on high-risk environments (e.g., autonomous driving accidents or medical errors) or simple comparisons between humans and AI. These studies reveal that negative events often elicit strong responsibility attribution reactions. However, there is limited research on multi-responsibility attribution scenarios involving human-AI collaboration, particularly in low-risk, everyday applications.

Proposed Solution

• Proposed Methods or Solutions:
  The authors conducted responsibility attribution experiments using semi-virtual scenarios, integrating psychological attribution theory, attention bias, and the asymmetry of blame and praise.

• Innovations:
  By simulating everyday AI systems (defined as task-assisting AI for non-life-threatening tasks that do not require professional training), the study breaks away from the traditional focus on high-risk scenarios and explores attribution differences in positive and negative outcomes.

• Implementation Steps and Key Techniques:

  1. Designed scenarios involving eight everyday tasks (e.g., time management, education, fitness) with variables such as perspective (first-person vs. third-person) and outcome (positive vs. negative).
  2. Recruited 321 participants to evaluate scenario outcomes (good or bad) and quantify responsibility attribution between humans and AI.
  3. Applied Bayesian statistical models to analyze quantitative data and conducted reflexive thematic analysis to interpret qualitative data on attribution mechanisms.

Research Findings

• Specific Findings:

  1. Participants tended to assign greater responsibility to AI systems, regardless of observer or executor perspective.
  2. For positive outcomes, responsibility was more likely to be jointly attributed to both humans and AI, while for negative outcomes, responsibility shifted more toward a single entity (without consistent attribution to either humans or AI).
  3. Participants perceived AI systems as praiseworthy for their insights and capabilities in positive outcomes but criticized them for their limitations in negative outcomes.

• Advantages:
  The study uniquely reveals the complexity of responsibility attribution in everyday AI scenarios, contrasting with traditional research that suggests humans are more often viewed as responsible. It broadens the understanding of responsibility attribution in low-risk environments.

• Experimental or Evaluation Results:

  • Quantitative analysis showed that positive outcomes increased responsibility scores for both AI and humans, while negative outcomes tended to assign responsibility to a single entity.
  • Qualitative analysis indicated that participants viewed AI as a "responsible tool" rather than an autonomous agent.

• Limitations and Future Directions:

  • The virtual nature of the scenarios may have caused participants' emotions to misalign with real-world contexts. Future research should validate findings in real-world environments.
  • The authors suggest exploring how responsibility attribution evolves during long-term interactions with AI systems and designing more transparent systems to manage user expectations and responsibility.

Conclusion

This study significantly advances the theoretical understanding of responsibility attribution in human collaboration with everyday AI systems, emphasizing the differing impacts of positive and negative outcomes on attribution patterns. It also highlights why AI is perceived as a "responsible tool." These findings not only provide practical guidance for the design of future AI systems but also demonstrate notable differences in responsibility cognition between everyday AI and high-risk scenarios.